Kinetic Barcodes: Reading Several Cytokines From One Electrode With Cyclic Voltammetry

Measuring panels of cytokines, the small signaling proteins immune cells secrete, currently requires optical instruments: fluorescent bead barcodes read by a laser flow cytometer, or single-molecule counting in microwells. Electrochemistry is far cheaper and can even be printed on plastic, but electrochemical sensors usually give one number per electrode. We propose to put several cytokine assays on the same electrode and tell them apart by two properties of their redox reporters, the small molecules that give up or accept electrons when an assay has captured its target. The first property is the voltage at which each reporter reacts. The second is how quickly it exchanges electrons with the electrode. Cyclic voltammetry sweeps the voltage back and forth and records the current. Run fast enough, or with a large sinusoidal wobble, it records both properties, so reporters that react at nearly the same voltage can still be separated by speed. A decoder built on a physical model of the electrode then turns one recorded curve into several cytokine concentrations.

The practical stakes are clear in both fields. Printed wireless cyclic voltammetry tags already exist, but they read one electrode and one threshold (Ha et al., 2015). Clinically useful cytokine signatures involve several proteins at once, such as IP-10 with IL-1Ra after lung transplant (Chacon-Alberty et al., 2022) or CCL4 across nervous system inflammation (Lyu et al., 2024). More assays per electrode means fewer electrodes, contacts, and reagents on a disposable device. It also lets each channel carry an internal reference that corrects for electrode fouling, which is otherwise the main failure mode of voltammetric sensors in biofluids.

Core Claim

Electron-transfer rate can serve as a second barcode dimension, alongside formal potential, for reporters bound to the electrode surface. Formal potential is the midpoint voltage at which a reporter switches between its oxidized and reduced forms. For a reporter bound to the surface, the current follows

$$i(t) = nFA\,k^0\left[\Gamma_{red}\,e^{(1-\alpha)f(E(t)-E^{0\prime})} - \Gamma_{ox}\,e^{-\alpha f(E(t)-E^{0\prime})}\right],$$

where $n$ is the number of electrons transferred, $F$ is Faraday's constant, $A$ is the electrode area, $k^0$ is the standard electron-transfer rate, $\Gamma_{red}$ and $\Gamma_{ox}$ are the surface amounts of reduced and oxidized reporter, $\alpha$ is the transfer coefficient, $f = F/RT$, and $E(t)$ is the applied voltage waveform. Each assay's signal scales with its surface amount $\Gamma$, which is set by how much cytokine was captured. Its shape is set by the pair $(E^{0\prime}, k^0)$.

Pulse methods such as differential pulse voltammetry reduce each reporter mostly to a peak height at one voltage, so $k^0$ is largely thrown away. Fast cyclic sweeps and the higher harmonics of large-amplitude sinusoidal voltammetry depend strongly on $k^0$. Bell et al. (2010) derived analytical expressions for exactly this surface-bound case using the protein azurin. They showed that the 2nd and 3rd harmonics carry the electron-transfer current while the capacitive charging background stays mainly in the fundamental.

Our hypothesis is this. Reporters whose voltages overlap within the 50–100 mV range, where ordinary peak deconvolution fails (Marenco et al., 2024), can still be quantified separately if their $k^0$ values differ by about an order of magnitude. Separating reporters by both voltage and speed should roughly double the number of assays a single electrode can support. This is the claim to test, not an established fact. We have found no paper that uses $k^0$ as a multiplexing axis, and none that performs multiplexed cytokine voltammetry at all.

Relation to Prior Work

Three bodies of work supply the pieces, and none puts them together. On the assay side, nELISA (Dagher et al., 2023) tethers each capture–detection antibody pair to its bead with DNA. It then uses toehold-mediated strand displacement, a DNA reaction in which a strand binding a short overhang pushes off another strand, to release a fluorescent oligo only when a full sandwich forms. The tethering reaches 191-plex without reagent cross-reactivity. We reuse this signal-generation logic. Our difference is where the signal goes: a redox reporter hybridizes to the electrode surface instead of a fluorophore appearing on a bead. The reporter's attachment chemistry and distance from the electrode are chosen to set its $k^0$.

On the readout side, Bell et al. provide the harmonic analysis for surface-bound redox species, and the grid-free integral solver of Lindsay et al. (2021) provides fast simulation of adsorption and diffusion during sweeps. It runs in 2.34 s versus 9.94–83.02 s for grid-based solvers. We use these as the forward model inside the decoder. Lindsay et al. assume equal diffusion coefficients and simple adsorption with no affinity binding. We need only the surface term plus a background from diffusing species, which fits within their assumptions.

On the decoding side, Al-Gharabli et al. (2019) showed that neural networks classify mixtures from cyclic square-wave voltammograms almost perfectly: ALSTM-FCN reached AUC 0.999 on their 11-class seawater set. They did not attempt concentration regression, however, and say so explicitly. Their result motivates using the full waveform. The finding that LDA was strong on their small 11-class explosives set cautions against assuming a large network is needed.

The constraining evidence comes from sensors that oxidize the analyte directly. Marenco et al. (2024) reached 75 nM for fentanyl and saw fouling after repeated biofluid exposure. That is why we use affinity capture with reporters rather than oxidizing the cytokines themselves, and why we build in a reference reporter.

Research Program

The work has three stages.

The first stage tests the physics. On gold and carbon electrodes we attach DNA-tethered reporters with established redox chemistry. We vary the reporter family to shift $E^{0\prime}$ and vary tether length and attachment position to shift $k^0$. For each construct we measure $(E^{0\prime}, k^0)$ by scan-rate series and by harmonic analysis. We then map which pairs a single electrode can quantify simultaneously when the reporter surface amounts are varied independently. The output is a "resolvable barcode map," and it decides whether the idea survives.

The second stage builds the assay. Each chosen barcode is attached to a DNA-tethered antibody pair modeled on the nELISA design, so that sandwich formation exposes a toehold that captures that reporter from solution onto the surface. One extra barcode sits on a strand that never responds to target. It serves as an in-channel reference, and the analyte signals are divided by it to correct for fouling and changes in electrode area. The starting panel of six cytokines is drawn from the cited clinical work: IP-10, IL-1Ra, CCL4/MIP-1β, IL-6, TNF-α, and IL-1β. IL-1β is included because its secretion is not predicted by its mRNA (Dagher et al., 2023).

The third stage builds the decoder. It fits the surface amount of every reporter by inverting the analytic surface model above, together with a background from diffusing species simulated with the Lindsay method. A small learned correction term absorbs non-ideal behavior such as spread in $\alpha$ and frequency dispersion. Fitted surface amounts are then converted to concentrations through per-analyte binding calibration curves. Because the decoder is built on the physics, drift and overlap show up as interpretable changes in fitted parameters rather than silent errors.

Baselines

The naive reference is spatial multiplexing: one electrode per cytokine with one reporter each, read by square-wave or differential pulse peak height against a standard curve. It uses the same antibodies, DNA constructs, electrodes, and sample volume split across channels. This is what an electrochemist would build by default. Our method beats it only if it matches its per-analyte limit of detection and accuracy while using a fraction of the electrodes. We will also report the regime where it wins outright, which is when electrode count and sample volume are unconstrained.

The strongest per-electrode competitor uses barcodes separated by voltage alone. Reporters are spaced as widely as the stable voltage window allows, read by differential pulse voltammetry, and separated by Gaussian peak deconvolution, following Marenco et al. (2024). It gets the same number of reporters per electrode and the same fitting budget. To beat it, we must support more analytes per electrode at a cross-talk ceiling of 10% of true signal.

The strongest data-driven decoder is the ALSTM-FCN architecture of Al-Gharabli et al. (2019), with its classification head replaced by a head that regresses multiple concentrations. It is trained on the same calibration voltammograms, with an equal hyperparameter search budget, and evaluated on held-out electrodes and days. The only numbers the file reports for it are for classification, such as AUC 0.999 on 11-SW. No regression result was found on Lacuna, so this comparison is new.

The optical reference is nELISA or bead cytometry run on split aliquots. It serves as ground truth for real samples. Its reported figures are the ceiling we do not expect to reach: 0.1 pg/mL sensitivity and a 1-plex versus 191-plex concordance of R² = 0.988 (Dagher et al., 2023).

Ablations isolate the key ingredient. The first gives all reporters the same $k^0$, which reduces the method to voltage-only barcodes with the same decoder. The second replaces the harmonic waveform with a DC sweep. The third removes the physical model from the decoder. The fourth removes the reference reporter.

Why not just…?

Why not just use one electrode per cytokine? This is the serious threat. For a six-plex benchtop panel it may simply be good enough. Our method earns its place only where contacts, area, or sample are scarce, as on printed single-channel devices like Ha et al. (2015), or where the reference reporter on the same surface measurably improves accuracy under fouling. The spatial baseline and the no-reference ablation settle this directly. If spatial arrays match us on accuracy in serum with the same total area, we will reframe the contribution as fouling-robust ratiometric channels rather than a gain in number of analytes.

Why not just use differential pulse or square-wave voltammetry, which suppress charging current better? For single-reporter quantification they likely will win, which is why they serve as the spatial baseline. They reduce each reporter to a peak position and height, discarding the $k^0$ information our barcode relies on. Bell et al. show that harmonics of a large-amplitude sinusoid remove most of the capacitive background while keeping kinetic information. The DC-sweep and equal-$k^0$ ablations show whether the kinetic dimension actually adds resolvable channels.

Why not just train a neural network on raw voltammograms? Al-Gharabli et al. show that networks separate mixtures well, but only for identity. Calibration data for multi-cytokine regression will be small, and their own LDA result suggests small data favors simpler models. The ALSTM-FCN regression baseline under a matched budget, tested on held-out electrodes and days, will show whether the physical model helps generalization. If it does not, we adopt the network.

Why not just keep optical bead assays? For discovery-scale plexity, nELISA is far ahead and we do not compete there. The electrochemical route targets small panels on disposable hardware. Transcript profiling is not a substitute either, since mRNA and secreted protein diverge (Dagher et al., 2023).

Proposed Experiments

No public benchmark for electrochemical cytokine multiplexing appears in the file, so we will build and release one. It has two parts. The first is a recombinant spike-in set in buffer and pooled serum, covering all six analytes on a factorial design of concentrations spanning 1 to 10⁴ pg/mL. That range is our target, not a reported capability. The second is a set of stimulated PBMC supernatants, echoing the design of the nELISA perturbation screen, measured in parallel by bead cytometry.

We report five metrics. Accuracy is per-analyte error in log concentration. Sensitivity is the limit of detection. Cross-talk is the signal that appears in one channel when only another analyte is spiked. Plexity cost follows Dagher et al.'s test: concordance between single-plex and full-plex assays on the same electrode. Stability is signal drift over repeated cycling in serum, with and without reference correction.

The decisive check comes first and is cheap. Two reporters with $E^{0\prime}$ within 30 mV of each other and $k^0$ differing by roughly 10× are placed on one electrode at independently varied surface amounts. If the harmonic decoder recovers both amounts with cross-talk under 10% across a 100-fold ratio, the kinetic barcode is real and the assay work proceeds. If not, the proposal falls back to voltage-only barcodes plus the reference reporter, and we report that negative result. The second decisive check is the full-panel concordance of single-electrode results against split-sample bead cytometry on PBMC supernatants, reported separately for IL-1β.